Which ChatGPT plan survives your Codex week?

Everything reduces to one unit, and here the unit is OpenAI's own: the credit. Since April 2026 Codex meters tokens rather than messages, and credits are what tokens are counted in. Sol, Terra and Luna convert into a Terra-equivalent agent-hour at their published token-price ratios, which are exact rather than blended: Sol is $5 in and $30 out per million against Terra's $2 and $12, so a Sol hour counts as 2.5. Luna is $0.20 and $1.20, so a Luna hour counts as 0.1. The whole family is a clean 25 to 10 to 1.

• Published. The plan ladder: Free $0, Go $8, Plus $20, Pro from $100 for 5x higher rate limits than Plus and $200 for 20x, Business $20 per user per month on annual billing or $25 monthly with a two-seat minimum, Enterprise and Edu on request. Read from OpenAI's Codex pricing documentation on 20 August 2026. Note that this supersedes the single $200 Pro rung described in older writing, including our own guide: the $100 Pro tier arrived in April 2026.
• Published. The two meters. OpenAI's wording is that local messages and cloud chats share a rolling five-hour window, and that additional weekly limits may apply. Every Codex surface counts against the same allowance: the CLI, the IDE extension, the desktop app, Codex cloud and iOS. Only an API key sits outside both.
• Published. Messages per five-hour window, which is the number this tool converts into capacity. On Plus: Sol 10 to 100, Terra 25 to 200, Luna 250 to 2,000. Pro 5x and Pro 20x are exactly five and twenty times each of those figures on every row, and a Business seat carries the same numbers as Plus. Free and Go publish no per-model row at all and the Codex model documentation lists Terra only on both, which is why a Sol share breaks them in the ranking above rather than merely costing more.
• Published. The credit rate card, per million tokens: Sol 125 in and 750 out, Terra 50 and 300, Luna 5 and 30.
• Published. API list rates per million tokens, short-context tier: Sol $5.00 in, $0.50 cached, $30.00 out; Terra $2.00, $0.20, $12.00; Luna $0.20, $0.02, $1.20. Long context is a separate column at 2x input and 1.5x output across the whole family. Read from OpenAI's API pricing on 20 August 2026.
• Derived, and checked. A credit is $0.04. OpenAI publishes the rate card in credits and the price list in dollars but never states the exchange rate, so we solved for it: Sol's 125 credits per million input tokens against its $5.00 per million input tokens gives four cents, and the same four cents reproduces all six published figures across the three models to the cent. The practical consequence is worth stating plainly: Codex credits are sold at exact API parity, so buying overflow is never more expensive per token than an API key.
• Derived, and checked three ways. The five-hour window capacity. OpenAI publishes message ranges rather than a budget, so we inverted them against its own note that GPT-5.6 usage averages 5 to 40 credits per message. On Terra both ends close exactly on the same figure: 25 expensive messages at 40 credits and 200 cheap ones at 5 credits are both 1,000 credits. That same 1,000-credit budget then reproduces Luna's published 250 to 2,000 range exactly at both ends, and Sol's low end of 10 exactly. Only Sol's published high end is about 25 percent more generous than the budget implies. So Plus and a Business seat carry 1,000 credits per window, Pro 5x carries 5,000, and Pro 20x carries 20,000, which is $40, $200 and $800 of inference respectively.
• Derived, with no published basis at all. Free at 50 credits per window and Go at 125. OpenAI gives no per-model row for either and describes them only as quick and lightweight coding tasks, so these are our floor estimates at 5 and 12.5 percent of Plus. Treat both rows as directional. What is not a guess is the model gate: neither plan carries Sol.
• Derived, and the weakest number on this page. The weekly limit. This is the one figure OpenAI genuinely does not publish: its wording is that additional weekly limits may apply, with no number attached. We set the weekly ceiling at three windows' worth of credits, so 3,000 on Plus, 15,000 on Pro 5x and 60,000 on Pro 20x. That is calibrated against the widely reported experience that an account running Codex most of the working day meets a weekly limit in the second or third day, and against the fact that a cap OpenAI thought worth warning about must bind somewhere well below the 33 windows a week contains. If you only take one number here with a pinch of salt, take this one.
• Derived, and the reason the ranking is not sharper than its inputs. Because that weekly figure is an estimate, a plan has to fit inside 85 percent of it before this tool calls it a survivor. The margin applies to the weekly ceiling only, not to the five-hour window, which is derived but checked three ways against published numbers. Without it a profile landing four percent under an estimated cap would outrank a plan with real headroom, which is advice built entirely out of the weakest figure on the page. It is the same instinct behind our Claude tool using the low end of Anthropic's published ranges: a plan you pick on the optimistic end is a plan that fails in week two.
• Derived. The heavy-context factor of 1.43. Unlike Anthropic, OpenAI publishes the surcharge itself: past the long-context threshold a request bills at 2x input and 1.5x output, and on the token shape below that makes a crossing request cost 1.86x. What it does not publish is how often an agent session crosses, so the toggle assumes about half of requests do. An agent that reads twelve files before editing one crosses far more easily than a chat does.
• Derived. Token throughput of about 2.7 million tokens per agent-hour with a warm prompt cache, which is what turns hours into credits and prices the API row. It comes from a worked twenty-turn agent session (115k cache writes, 1.24M cache reads, 16k output) taken at two sessions an hour, and on Terra's rates it lands at $1.34, or 33.5 credits, per Terra-equivalent agent-hour.
• Ours, and in the ranking on the same rule as everything else. Continuum's hosted plans carry a weekly dollar allowance consumed at the same API-equivalent burn as every other row, and fit only while the week's dollar figure stays under the allowance. Two guards keep it honest: a Continuum plan can never displace a cheaper OpenAI plan that also survives, and on an exact price tie the OpenAI plan wins. What the allowance buys is hosted inference on any model through an OpenAI- and Anthropic-compatible endpoint, so it is priced here at OpenAI list rates to keep the comparison like for like; Continuum's own per-model rates are on the pricing page and are not identical, so read the headroom figure as indicative.
• Ranking rule. The metered portion of a row has to come in about 25 percent under a fixed monthly price before it is allowed to outrank it. Inside that band the known invoice wins, because the estimate's own variance is wider than the gap. Only the metered part carries the penalty: a $100 plan with $12 of credits on top is mostly a known invoice, and taxing all of it would price a certainty as a guess.
• Judgement call. Meeting the five-hour window is treated as a degraded week, not a broken one, which is the opposite of how our Claude tool treats it and the difference is real. The window refills continuously, and Luna costs a tenth of Terra on every paid plan, so an overflowing window costs you an afternoon or one /model command. The weekly limit is the hard stop, because nothing you switch to clears it.
• Simplification. Cap-fire days assume your week is spread evenly across the days you selected. A real Monday-heavy week fires earlier than the number shown.
• Simplification. The credit rows price the larger of two shortfalls, weekly and window. Window overflow is charged per window rather than once a week, since it recurs every time the window refills.
• Not modelled. Fast mode, which is a priority service tier that bills credits at a higher rate on the models that support it. Enterprise and Edu, which have no published rate limits and scale with workspace credits. Annual billing on personal plans. And a reported temporary suspension of the five-hour restriction on some paid plans in July 2026: OpenAI's own pricing documentation still describes the window as live today, so this tool models it as live.

Prices, limits and rate cards read from OpenAI on 20 August 2026. OpenAI has changed Codex pricing twice this year; check the numbers before you commit a year to one.